1. King : Estimating latency between arbitrary Internet end hosts Krishna Gummadi, Stefan Saroiu Steven D. Gribble University of Washington Presented by: Amit Mondal

2. A Question: What can we do with a tool that could estimate latency between any two arbitrary hosts accurately? A B C C can measure latency between A and B

3. Potential uses of such a tool  Building topologically sensitive overlays  Selecting a close replica server  Scaling wide-area measurement studies involving latency estimation Detour etc., current state of the art techniques allow at most a few hundred end hosts to be measured

4. Current state of the art  Use techniques like IDMaps and GNP inaccuracy in estimates: We need a tool that can measure latency rather than compute it issues with deployment: IDMaps requires additional infrastructure;  Use shared measurement infrastructure e.g., trace-route servers, PlanetLab,

5. King: a new measurement tool  Estimate latency between arbitrary end hosts  Requires no additional infrastructure leverage existing DNS infrastructure enabling a large fraction of Internet hosts to be measured  Provides highly accurate latency estimates  Fast and light-weight requires only a few DNS queries per estimate

6. Outline  Motivation  How King works  Evaluation of King  Conclusions

7. Name Server near Host A Name Server near Host B Host BHost A Actual Latency Between End Hosts Latency Estimated By King How King Works: The Basic Idea  Challenge 1: How to find name servers that are close to end hosts  Challenge 2: How to estimate latency between two name servers

8. Challenge 2: How do we estimate the latency between name servers? Our Client C (King) Name Server B foo.bar Name Server A 1. Request Q: Resolve xyz.foo.bar 4. Reply Q (Forwarded) 2. Request Q (Forwarded) 3. Reply Q: IP addr of xyz.foo.bar

9. Success of Recursive DNS  For King to work, name servers must support recursive queries in a large random sample, > 75% of name servers supported recursion translates to > 90% success rate given a pair as we can measure from A->B, or B->A

10. Challenge 1: How to find DNS servers nearby the end hosts  Assumption: Authoritative name servers for the host are closeby (topologically and geographically)  This assumption may not always hold, but our evaluation shows that it is true in general e.g., AOL is an exception  To find an authoritative name server given host name, use forward name resolution given host IP, use reverse lookup in in-addr.arpa domain

11. Selecting a close name server  When multiple authoritative name servers are deployed, how do we choose a close one? select the server with longest matching DNS suffix and IP prefix with end host

12. Outline  Motivation  How King works  Evaluation of King  Conclusions

13. Evaluation of King  How accurate is King?  What are the causes of inaccuracy?  Can King identify its own inaccurate estimates?  Does King preserve order among its estimates?

14. Accuracy of King  Compare the accuracy of King with IDMaps  Methodology Measure true latency between 50 public Traceroute servers and 50 end hosts using Traceroute Estimate latency between the same endpoints using King and IDMaps Compare estimated latency with measured latency Metric used : Estimated Latency Measured Latency

15. Accuracy of King  King is far more accurate than IDMaps  King tends to under-estimate latencies typically, name servers have higher BW and lower latency last hop links than end hosts

16. Evaluation of King  How accurate is King?  What are the causes of inaccuracy?  Can King identify its own inaccurate estimates?  Does King preserve order among its estimates?

17. Causes of Inaccuracy in King  Authoritative name servers may not be close to end hosts  Latency estimation between the name servers might be inaccurate application level latency at DNS servers to resolve the query

18. Are authoritative name servers close to their end hosts?  In a random sample, 70-80% of end hosts and their name servers are separated by less than 10-20 msec  Our conclusion contradicts earlier studies !!  Possible explanations: We looked at more metrics divergent path hop count – a misleading metric used primarily in other studies; divergent path latency – tells a different story Unknown bias in our random samples

19. Application level latency for DNS servers  Methodology: selected a large number sample of name servers measured latency to servers using Ping and DNSPing (iterative DNS query) over time Query resolution latency = DNSPing – Ping  Application level latency negligible Implication: King estimates between name servers are very highly accurate

20. Evaluation of King  How accurate is King?  What are the causes of inaccuracy?  Can King identify its own inaccurate estimates?  Does King preserve order among its estimates?

21. Can King identify its own inaccurate estimates?  Primary cause of error in King authoritative name servers far from their end host  Simple heuristics based on the lengths of DNS suffix and IP prefix match work well

22. Evaluation of King  How accurate is King?  What are the causes of inaccuracy?  Can King identify its own inaccurate estimates?  Does King preserve order among its estimates?

23. Does King preserve order among its estimates?  Sometimes preserving order among estimates is more important than their accuracy Applications like server selection  King does very well at preserving order among its estimates very high correlation coefficient (>0.8) between the orderings of estimated and true latencies large latency last hops do not effect order

24. Summary of evaluation  King is far more accurate than IDMaps King errors more when it under-estimates due to large last hop latencies of end hosts accuracy of estimates between name servers is even higher  The primary cause of error is the authoritative name servers that are far from their end hosts King uses heuristics to identify such errors  King preserves excellent order among its estimates

25. Validating King’s utility for wide-area measurement studies  The Detour study showed that default routes are inefficient, and alternate routes can have better latency. they were limited to 35x35 data points  We repeated study using King we gathered 193x193 data points The data points were name servers chosen using King’s self- evaluation heuristics it took less than 4-5 hours using a single machine our results were consistent with those from earlier study

26. Conclusions  We presented King; a new measurement tool that can estimate latency between arbitrary Internet end hosts does not require any additional infrastructure as it leverages existing DNS infrastructure fast and light-weight  Our evaluation of King confirms that it is accurate it preserves order among its estimates  We showed that King can be used in scaling wide- area measurement studies

27. Thank you!